Interactive Loss Landscape Exploration (ILLE): Direct Visualisation of Neural Network Loss Surfaces via Forward Pass Rendering and On-Demand Navigation

We introduce Interactive Loss Landscape Exploration (ILLE), a visual optimisation paradigm that renders the neural network loss surface directly from forward pass evaluations — with no backpropagation, no gradient computation, and no surrogate approximation. The central insight is that the loss function L(W) can be evaluated at any weight configuration W via a single forward pass, and that systematically evaluating L across a 2D grid of weight values produces a directly renderable topological map of the loss landscape. Local and global minima appear as valleys on this surface and can be identified visually and algorithmically, transforming weight-space navigation from a blind sequential search into an interactive visual task. For computational efficiency, ILLE renders on demand: a coarse overview is computed first, and high-resolution forward passes are fired only as the user navigates into a region of interest, analogous to tile loading in a digital mapping application. We validate ILLE on a two-weight sine-curve regression task across eight grid resolutions (100 to 40,000 forward passes). At resolution 200 by 200, the global minimum is recovered at w = (1.0151, 0.0101) against the analytical solution (1.0000, 0.0000), achieving 99.57 percent accuracy with total wall-clock time 0.438 seconds. Minimum identification converges at resolution 20 by 20 (400 passes, 6ms), with additional resolution providing finer localisation only. The framework is model-agnostic and extends to deeper networks via weight-space projection.